This PR is to provide a performant example of training a fairly large model (ResNet-50) on a large dataset (ImageNet) that demonstrates the MxNet backend perf for a single GPU and scaling to multiple GPUs (tested up to 8 P100s).
I'm sure there are things that may need to be iterated on, especially the need to copy over data.py from MxNet, since it's beyond the scope of the root of the mxnet Python package (it's from mxnet/example/image-classification/common, but the root of the Python MxNet package starts at mxnet/python/mxnet, so data.py cannot be imported from keras/examples). I'll make the necessary changes as I get feedback, of course.
This PR is to provide a performant example of training a fairly large model (ResNet-50) on a large dataset (ImageNet) that demonstrates the MxNet backend perf for a single GPU and scaling to multiple GPUs (tested up to 8 P100s).
I'm sure there are things that may need to be iterated on, especially the need to copy over data.py from MxNet, since it's beyond the scope of the root of the mxnet Python package (it's from mxnet/example/image-classification/common, but the root of the Python MxNet package starts at mxnet/python/mxnet, so data.py cannot be imported from keras/examples). I'll make the necessary changes as I get feedback, of course.